Enterprise consolidates 200 apps onto single self-hosted LLM via multi-axis expert merging

The paper doesn't just show that you can merge specialized models; it demonstrates that production bottleneck data can drive post-training without creating reward conflicts across domains. That's the constraint most enterprises haven't solved yet.

This consolidation pattern mirrors what we've seen in professional services and agriculture. Gilbert + Tobin's governance framework and John Deere's domain-specific embedding both faced the same underlying tension: regulatory or data-residency requirements force on-premise hosting, but that infrastructure cost pressures companies to unify rather than proliferate. The difference here is methodological: instead of deploying a single general model (which underperforms on specialized tasks), this approach uses production traffic to identify which axes matter most, then trains targeted experts and merges them. It's the missing link between the operational necessity we documented and the technical execution. The document VLM paper from the same day also addresses this (cost-quality parity through specialized fine-tuning on production data), but that's a single-task system; this one handles 200+ applications.

If other enterprises publish similar consolidation results within the next 12 months using the same GRPO-plus-SLERP approach, that signals the method is reproducible and will likely become standard practice. If they don't, the bottleneck is probably in identifying which production signals actually matter for their specific application mix, not in the merging technique itself.